Yiak Wang Yi

[Yiak Wang Yi @YiakYi]

@Barazany @SemiAnalysis_ it means improvement of quality of silicon chip production, hence less usage of inspection tools.

[Yiak Wang Yi @YiakYi]

Check out my latest article: Buidling Native Agent Driven Workflow https://t.co/d20rl1czy6 via @LinkedIn

[Yiak Wang Yi @YiakYi]

7.31:
"we plan to open up the model weights in the coming days, subject to applicable laws and regulations. Hardware compatibility has been a key consideration since the earliest stages of H3's design."

8.3 : production ready video Gen M is open 🤗😍

https://t.co/kVs5xOH70d

[LEI WANG @yiakwy2023]

@Francis_YAO_ Linear attention is amazing . Now with nope MLA linear attention has fundamental transformed the logics of KV cache D. Your kv cache will never scale with sequence length and work very well. So next question , where is K3 flash ?

[LEI WANG @yiakwy2023]

Official K3 1 bit weights are out in @huggingface @UnslothAI 🔥🔥🔥

https://t.co/0xqpIUAq8A

[Yiak Wang Yi @YiakYi]

@UnslothAI We are also working on this problem. Thank you for the contribution!

[LEI WANG @yiakwy2023]

quick things you should know about CDNA5 https://t.co/XY1PftvG9B for Mi455:
1. WarpSIze : 32 , no more 64 （4 cycles x 16 SIMD to 1x cycles x 32 SIMD)
2. WGP （2 CU）for multicast
3. New Tensor Data Mover analogue to TMA in Hopper
4. 1024 vgpr for simd
5. Better Noc cross XCU

[Yiak Wang Yi @YiakYi]

@geekbb For more details of the insight see the following:

RedNote : https://t.co/XbZXFvf4d7
BlueBook : https://t.co/oM1v8u5XfS
X : https://t.co/m8u2Pr0pT7

[Yiak Wang Yi @YiakYi]

For more details of the insight see the following:

RedNote : https://t.co/XbZXFvf4d7
BlueBook : https://t.co/oM1v8u5XfS
X : https://t.co/m8u2Pr0pT7

[Yiak Wang Yi @YiakYi]

"Apparently K3 linear attention has fundamentally transformed logics of SSD and KV cache, you cannot assume that SSD grows with KV cache", an investor visited me today in campus this morning.

[Yiak Wang Yi @YiakYi]

"Apparently K3 linear attention has fundamentally transformed logics of SSD and KV cache, you cannot assume that SSD grows with KV cache", an investor visited me today in campus this morning.

[Yiak Wang Yi @YiakYi]

Deepseek V4 flash (7-31) finally out ! And after few days thinking about K3 KDA and AttnRes I just found some new opportunity where we can apply optimization upon.

[LEI WANG @yiakwy2023]

Deepseek V4 flash （7-31）delivered and out performs GLM 5.2！

[Yiak Wang Yi @YiakYi]

@suno For more technical details see :

For more info , see this article : https://t.co/NzWTtnqvuL

[1] Qwen-Music Report: https://t.co/A9tnD2y5h9
[2] BTC-LLM (our lab's quantization work): https://t.co/Nfdx7nYnRW
[3] StreamFlow (NeurIPS 2025): https://t.co/36hrZ6LafA

[Yiak Wang Yi @YiakYi]

recently dived into Qwen-Music [1], a new open music generation system that achieves SOTA results in expert evaluations—outperforming Suno V5.5 and Mureka V8 on key metrics, and ranking 3rd on the Artificial Analysis leaderboard (as JazzCat).

[Yiak Wang Yi @YiakYi]

A colleague from our research group contributed to this work, so to better guide our subsequent audio optimization efforts, I analyzed its core computational characteristics.

The work now is very close to the top AI music Suno V5.5 @suno

[Yiak Wang Yi @YiakYi]

@a_karvonen Meta Monarch should be pushing for week 0 support. The recently released models like Think Machines and K3 are worth considering.

Monarch's support has been quite disappointing. I'm not sure if the management team even understands the critical priority.

@marius @joespeez

[Yiak Wang Yi @YiakYi]

@a_karvonen @marius @joespeez This allows teams that are eager to deliver quickly to get immediate results on top of our work. It's a win-win.

On the other hand, Meta's solution was to develop TorchForge to address usability issues, but that ended up diverting their team's focus and attention.

[Yiak Wang Yi @YiakYi]

@a_karvonen @marius @joespeez A few observations:

Most researchers prefer to modify code within their familiar frameworks. New lightweight frameworks mainly attract fresh graduates as early adopters.

Drawing from our experience with Miles last time, I believe the key to success is providing week 0 support.

[Yiak Wang Yi @YiakYi]

@MiaAI_lab hi we have enabled deepseek-v4 q2/q4 in a single dgx, that means with 2 spark you can implement P1D1 disaggregation to get even better performance. Dgx has maximum 300 toks/sec prefill speed, and 16 tokens decoding speed . Did you try it ?

[LEI WANG @yiakwy2023]

Your Muon Optimizer can be even more faster !

Accelerating SymGemm with Triangular Scheduler (friendly for NoC clusters and L2 cache affinity!)

## Intro

The author from MIT Laker New House once delivered a partial implementation of accelerated symmetric matrix multiplication